Job description

Description Smeetz is an AI-powered unified commerce platform for the European leisure and entertainment industry.
Theme parks, family entertainment centres, museums, zoos, and theatres run their ticketing, seating, F&B, and payments on us.
This role owns our two AI surfaces: the Smeetz MCP server and the AI chatbot.
You own them end to end.
You talk to the venues using them, you decide what to build next, you write the production code, and you watch what happens once it is live.
There is no product manager between you and that decision.
Reports to the VP of Engineering.
No direct reports.
The authority in this role is over a surface area, not over people.
Key responsibilities A product engineer owns a piece of the product end to end: choosing what to build, building it, and improving it once real customers are using it.
Here that piece is our AI surfaces.
Own the Smeetz MCP server and the AI chatbot.
Both are yours end to end: the tool surface, the agent behaviour, and the guardrails on what they are allowed to do.
Decide what to build next.
Run discovery on your own area.
Talk to venue operators, read the traces, pick the work, write the spec yourself, ship it, then measure it.
Own reliability and performance.
Latency, error rates, and the call patterns clients actually generate.
You get full observability data to work from and the mandate to act on it.
Build the eval layer.
Make prompt and model changes measurable before they ship, not after a customer notices.
Write things down.
The surfaces you own should be understandable by the next engineer without you in the room.
Documentation sits in your objectives, not in the gap at the end of a sprint.
Raise the bar on AI-native delivery.
Extend our shared Claude Code skill library so the rest of engineering ships the way you do.
A published dual career ladder.
Grow to Staff AI Product Engineer without becoming a manager, on pay bands comparable to the management track.
Full ownership of two AI surfaces that are core product, not a side project.
Direct contact with the venue operators who use what you build.
Modern AI-native tooling: Claude Code with a shared internal skill library.
Fully remote position.
Competitive salary based on experience.
Band to confirm.
Professional requirements Minimum 5 years shipping production code, including at least one system you owned end to end with the on-call that comes with it.
Minimum 2 years building LLM features that reached real users: tool calling, structured outputs, evals, cost and latency control, and the failure modes that only appear in production.
Prototypes do not count.
Demonstrated expertise in an AI-native SDLC.
You work in Claude Code or an equivalent agentic environment every day, and you build your own skills, hooks, and automations rather than using AI as autocomplete.
Be ready to show what you automated and what it replaced.
Experience designing an API or tool surface for agents rather than for humans: naming, granularity, and error messages a model can recover from without a human stepping in.
You run discovery yourself.
Bring customer conversations you led, specs you wrote, and success metrics you set and were measured against.
Fluent in English.
Additional preferred skills Langfuse, or another LLM observability tool such as LangSmith, Braintrust, or Phoenix.
Ours is Langfuse and it is already wired into the MCP server.
The MCP specification itself, beyond having used it as a client.
The Anthropic Claude API and tool use.
Most of what we build sits on Claude.
Retrieval and evals over a support knowledge base.
Intercom Fin experience is directly relevant.
Guardrails for agents with write access: permissions, confirmation steps, prompt injection, PII.
French or German.
Personal attributes Owner: Defaults to ownership.
The surface works because you are in the seat.
AI-native: Lives inside Claude and agents.
Treats AI as leverage, not as a side project.
Product-minded: Starts from the operator's problem, not from the ticket.
Decisive: Kills their own ideas and can say what they gave up and why.
Writer: Leaves the codebase better documented than they found it.
At ease in ambiguity: There is no groomed backlog for this area.
That is the appeal, not the problem.
Measured: Defines success before building and accepts the number.